Cash management models determine policies based either on the statistical properties of daily cash flow or on forecasts. Usual assumptions on the statistical properties of daily cash flow include normality, independence and stationarity. Surprisingly, little empirical evidence confirming these assumptions has been provided. In this work, we provide a comprehensive study on 54 real-world daily cash flow data sets, which we also make publicly available. Apart from the previous assumptions, we also consider linearity, meaning that cash flow is proportional to a particular explanatory variable, and we propose a new cross-validated test for time series non-linearity. We further analyze the implications of all aforementioned assumptions for forecasting, showing that: (i) the usual assumption of normality, independence and stationarity hardly appear; (ii) non-linearity is often relevant for forecasting; and (iii) common data transformations such as outlier treatment and Box-Cox have little impact on linearity and normality. Our results highlight the utility of non-linear models as a justifiable alternative for time series forecasting.
↧